A norm that used to be enforced one manager at a time
Massoud Mansouri, a consultant occupational physician, has published an editorial in Occupational Medicine on artificial intelligence and neurodiversity at work. It’s three pages, and it’s a call for occupational health to pay attention rather than a body of new evidence — but it assembles a set of facts that deserve a wider audience than its readership, and one of them is genuinely alarming.
That one first. Research published in Autism Research has measured the prevalence of bias against neurodivergence-related terms inside AI language models themselves. Not in how employers use them — in the models. The associations are already there, learned from text written by people, sitting inside systems now being asked to evaluate people.
Around that sit two more layers. Recruitment: algorithmic systems increasingly filter on behavioural and communication patterns derived from large datasets, and those patterns reflect common ways of interacting by construction. Where such systems act as early-stage filters, difference can determine the outcome before any human judgement is applied. Then workforce analytics: systems inferring concerns from patterns of absence, communication or productivity, where — as Mansouri puts it — patterns that appear atypical may reflect a different working style rather than a difficulty, and a system without that context will misread them.
The shift worth naming is one of scale and location. The neurotypical norm has always been enforced in workplaces, but it was enforced retail: one manager’s raised eyebrow at a blunt email, one panel’s discomfort with someone who didn’t make eye contact, one appraisal that marked down a person whose productivity came in bursts. Each instance was contestable, appealable, sometimes just avoidable by finding a better manager. What is happening now is that the same norm is being trained into infrastructure — applied uniformly, at volume, before a human is involved, by a system that cannot be asked why. Bias moves from being a thing some people do to being a property of the pipeline.
Masking, demanded by a system that never explains itself
Mansouri identifies masking as the mechanism through which this reaches the person, and he’s right, but the connection is sharper when set against last week’s piece on psychological safety.
That argument ran: masking at work is not a coping strategy people happen to adopt, it is a measurement of the safety that isn’t there. You suppress how you actually work because being yourself would be penalised. The evidence Mansouri cites confirms the cost — workplace masking is associated with increased anxiety, depression and reduced wellbeing.
Now add a system calibrated to particular patterns of productivity, responsiveness and communication, evaluating continuously. The pressure to adapt no longer comes from a person who might, with effort, be educated. It comes from something that defines acceptable performance in advance, applies it to everyone, gives no reasons, and hears no appeal. Mansouri’s phrase for the result is sustained pressure to adapt in order to meet system-defined expectations — which is masking, made permanent and unarguable.
And there’s a compounding problem specific to automated evaluation: it removes the possibility of the accommodation ever being requested. A human manager can be told that your productivity is non-linear, that you need written instructions, that the video interview is measuring your discomfort rather than your competence. An early-stage algorithmic filter cannot be told anything. The candidate is rejected before disclosure could occur, before the duty to make reasonable adjustments is triggered, and — because the system produces no explanation — without anyone, including the employer, knowing why. The screening happens in a place where the legal protections were designed to apply, and where they now arrive too late.
The same technology, running the other way
The honest account has to hold the other half, and this is the part I’d argue more strongly than the editorial does.
Mansouri notes the assistive potential — tools supporting organisation, summarisation and communication, reducing cognitive load and improving accessibility, with a systematic review finding benefits for adaptive functioning and task management in neurodevelopmental conditions. He’s careful that most of that work was done outside workplaces, and the mechanisms are inferred to be transferable rather than demonstrated. He also notes that flexible working, which these systems can enable at scale, measurably narrows employment disparities between neurodivergent and neurotypical workers.
What that undersells is what the evidence from this corpus already shows. The twice-exceptional university students interviewed earlier this month named Claude and ChatGPT unprompted, alongside text readers and project management software, as part of a self-built support apparatus they had assembled because the institution’s provision fitted neither of their exceptionalities. Nobody prescribed it. They found it, because it did something for them that the formal system didn’t: structured a task, held the thread, translated between how they think and what the assessment demanded.
That is cognitive scaffolding, and I’ve written before about what it means to lean on it — that the question worth asking is not whether the support is external, since all executive support is external, but whether it extends capacity or replaces it. The relevant point here is simply that both things are true simultaneously. The same broad technology is being used to screen people out and to help them work. It is not one thing with a valence.
The variable, in every case, is what the system was built to do. An AI configured to score candidates for fit against a normative interaction pattern will exclude difference, reliably and at scale, because that is the task it was given. An AI configured to structure a task, summarise a meeting, or translate a brief into a format someone can act on will extend capacity, for the same reason. The technology has no preference. The specification does — and the specification is written by people who mostly haven’t been asked to think about who their normal was drawn from.
What the thing is calibrated to measure
Which is why the governance question is the practical one, and where the editorial is genuinely useful.
Mansouri notes that neurodevelopmental conditions may meet the definition of disability under the Equality Act 2010, and that where AI-enabled recruitment or performance processes disadvantage such individuals, this may raise issues of indirect discrimination and the duty to make reasonable adjustments. In the UK, oversight is distributed across health and safety, equality, employment and data protection law rather than any single AI statute, with government guidance on responsible AI in recruitment sitting alongside it. The EU has gone further, with the AI Act classing certain employment-related systems as high-risk and requiring additional safeguards.
His conclusion — that automated processes should not be assumed neutral, and that occupational health has a role in evaluating whether these systems increase psychosocial demands or read cognitive variability as poor performance — is correct and modest. I’d put the practical version more bluntly for anyone deploying these tools: you are already accountable for what your systems do to disabled candidates and staff, whether or not you know what they do. “The model decided” is not a defence that exists in law.
The deeper point is the one my corpus has made about every institution it has examined this year, arriving now in a new medium. A system that measures conformity to a norm will find neurodivergent people wanting, and it will do so with the appearance of objectivity, because a number came out of a machine rather than out of someone’s discomfort. A system built to extend what people can do will find the same people capable. Neither result is a discovery about the people. Both are readouts of what the builder chose to measure — and the only difference between the gatekeeper and the scaffold is which question got specified before anyone pressed run.
Citations
Mansouri, M. (2026) — Neurodiversity and artificial intelligence: psychological risk and potential for inclusion — Occupational Medicine
Brandsen, S., Chandrasekhar, T., Franz, L. et al. (2024) — Prevalence of bias against neurodivergence-related terms in artificial intelligence language models — Autism Research
Pryke-Hobbes, A., Davies, J., Heasman, B. et al. (2023) — The workplace masking experiences of autistic, non-autistic neurodivergent and neurotypical adults in the UK — PLOS ONE
Davies, J., Heasman, B., Livesey, A. et al. (2023) — Access to employment: a comparison of autistic, neurodivergent and neurotypical adults’ experiences of hiring processes in the United Kingdom — Autism
Branicki, L., Brammer, S., Brosnan, M. et al. (2024) — Employment outcomes of neurodivergent individuals and flexible working — Human Resource Management
